Loading lesson…
AI Research Foundations / Method
A three-pass method for reading efficiently and critically.
Reviewed by Yuvaraj
A research paper is not a novel, and reading it front-to-back at a constant pace is the slowest possible way to extract its value. Skilled researchers read strategically: they spend a few minutes deciding whether a paper is even worth their time, and reserve deep effort for the rare paper that rewards it. This lesson teaches the three-pass method, a disciplined way to read that scales your attention to a paper's actual relevance, and the critical questions that separate a real result from a well-dressed claim.
Most people treat reading as passive absorption: eyes move across text, and understanding is supposed to accumulate on its own. It does not. A dense eight-page conference paper can encode months of work, and its authors have compressed it aggressively, every figure, every ablation row, every hedged sentence in the limitations section is load-bearing. Reading well means interrogating the paper: forming expectations, checking them against the evidence, and noticing what the authors chose not to say.
The single biggest efficiency win is refusing to read most papers deeply. You will encounter far more papers than you can absorb. The three-pass method, popularized by S. Keshav, gives you a way to triage: a cheap first pass tells you whether to invest a more expensive second or third.
Ask about this lesson, or about anything in AI. Answers cite the lessons they draw on.
Finished this lesson?
Mark it complete to earn XP, keep your streak, and schedule a review.
The passes are cumulative and self-limiting. After Pass 1 you can usually put the paper down. After Pass 2 you understand the contribution well enough to summarize it to a colleague or cite it responsibly. Only Pass 3 requires the multi-hour commitment, and you should perform it on a small fraction of what you read.
Pass 1 answers five questions: what category of paper is this (a new method, an analysis, a survey), what problem does it address, what is the broad approach, are the assumptions plausible, and is it relevant to me? If the answer to the last is no, you stop, and you have lost only ten minutes.
Pass 2 is where you grasp the paper's thrust without drowning in machinery. Deliberately skip the proofs and the appendix derivations. Spend the saved effort on the empirical section: read table headers carefully, check what the axes of every plot mean, and ask whether the numbers actually support the sentence next to them.
Pass 3 is a virtual re-implementation. You are no longer a reader but a re-builder. This is the pass that surfaces unstated hyperparameters, convenient dataset choices, and gaps between what the method promises and what the experiments demonstrate.
Reading is search, not consumption
Think of a literature as a graph and the reference list as its edges. Pass 1's glance at the references tells you where this paper sits in that graph, whose work it builds on and which community it belongs to. Following those edges backward (foundational citations) and forward (papers that later cite this one) is often more valuable than reading any single node exhaustively.
Reading comprehension is not the goal, judgment is. Once you understand what a paper says, you must decide whether to believe it. These questions are the core of that judgment, and Pass 2 and Pass 3 exist largely to answer them.
| Question | What you are really testing | Where to look |
|---|---|---|
| What exactly is claimed? | Whether the headline is broader than the evidence | Abstract vs. results; watch for words like "can" or "up to" |
| Compared against what baseline? | Whether the comparison is fair and well-tuned | Experimental setup, baseline descriptions |
| On what data? | Whether results generalize or overfit one benchmark | Datasets, splits, evaluation protocol |
| Would it reproduce? | Whether seeds, code, and hyperparameters are pinned | Appendix, released code, reproducibility statement |
| What limitations are under-stated? | Whether the authors hid the weak spots | Limitations section, and the experiments they omitted |
The last question is the most important and the most neglected. Authors are incentivized to present their work favorably; the interesting information is often in what is missing. If a method claims broad applicability but every experiment uses the same dataset, that gap is your finding, not theirs. If the baselines are all several years old, ask why the obvious recent competitor is absent.
A claim is a scope, not a slogan
"Our method improves accuracy" is not a claim you can evaluate. "Our method improves top-1 accuracy on ImageNet by 1.2 points over a tuned ResNet-50 baseline, at equal training compute" is. Rewrite every headline claim into this specific form before deciding whether the evidence supports it, the act of rewriting exposes exactly what the paper did and did not show.
For a stack of papers from a conference or a search, run Pass 1 on all of them first and sort into keep/discard. This batching keeps you in triage mode and resists the temptation to fall down a rabbit hole on the first interesting abstract. Only then schedule Pass 2 on the survivors, and mark at most one or two for Pass 3.
Keep notes as you go. A one-paragraph summary written in your own words after Pass 2, the problem, the idea, the evidence, and your single biggest doubt, is worth more than any highlighting. If you cannot write that paragraph, you have not finished Pass 2.
Common mistakes